Robust North Atlantic right whale detection using deep learning models for denoising.

Area of Science:

• Marine Biology
• Bioacoustics
• Artificial Intelligence

Background:

• Passive acoustic monitoring is crucial for North Atlantic right whale (Eubalaena glacialis) conservation.
• Noise from shipping and offshore activities severely degrades the accuracy of automated vocalization detection systems.
• Existing classifiers struggle with noisy data, leading to underestimation of whale presence.

Purpose of the Study:

• To develop a robust deep learning system for denoising North Atlantic right whale vocalizations.
• To improve the accuracy of right whale detection in acoustically challenging environments.
• To enhance the reliability of passive acoustic monitoring for conservation efforts.

Main Methods:

• Utilized deep learning, specifically denoising convolutional neural networks and denoising autoencoders, on spectrograms of audio signals.
• Implemented a novel classifier training approach incorporating residual signals from denoised audio.
• Simulated noisy conditions by adding various noise types (white, tanker, trawler, shot) at different signal-to-noise ratios.

Main Results:

• Denoising significantly improved detection accuracy across various noise types and signal-to-noise ratios.
• The vestigial-trained classifier demonstrated substantial performance gains.
• The system proved effective on previously unseen noisy recordings, outperforming baseline models.

Conclusions:

• Deep learning-based denoising is a highly effective method for improving right whale acoustic detection.
• The proposed system offers a robust solution for real-world passive acoustic monitoring challenges.
• This technology can enhance the accuracy and reliability of right whale population assessments and conservation strategies.